Fix deepseeed crash with Sentence Transformer Trainer (huggingface#1328)

Co-authored-by: ZhengHongming888 <hongming.zheng@intel.com>
Co-authored-by: Yaser Afshar <yaser.afshar@intel.com>

Author: 3 people

examples/sentence-transformers-training/nli/README.md

@@ -4,6 +4,8 @@ Given two sentences (premise and hypothesis), the task of Natural Language Infer
 
 The paper in [Conneau et al.](https://arxiv.org/abs/1705.02364) shows that NLI data can be quite useful when training Sentence Embedding methods. In [Sentence-BERT-Paper](https://arxiv.org/abs/1908.10084) NLI as a first fine-tuning step for sentence embedding methods has been used.
 
+# General Models
+
 ## Single-card Training
 
 To pre-train on the NLI task:
@@ -46,7 +48,29 @@ For multi-card training you can use the script of [gaudi_spawn.py](https://githu
 HABANA_VISIBLE_MODULES="2,3" python ../../gaudi_spawn.py --use_deepspeed --world_size 2 training_nli.py bert-base-uncased
 ```
 
-## Dataset
+
+# Large Models (intfloat/e5-mistral-7b-instruct)
+
+## Single-card Training with LoRA+gradient_checkpointing
+
+Pretraining the `intfloat/e5-mistral-7b-instruct` model requires approximately 130GB of memory, which exceeds the capacity of a single HPU (Gaudi 2 with 98GB memory). To address this, we can utilize LoRA and gradient checkpointing techniques to reduce the memory requirements, making it feasible to train the model on a single HPU.
+
+```bash
+python training_nli.py intfloat/e5-mistral-7b-instruct --peft --lora_target_module "q_proj" "k_proj" "v_proj" --learning_rate 1e-5
+```
+
+## Multi-card Training with Deepspeed Zero2/3
+
+Pretraining the `intfloat/e5-mistral-7b-instruct` model requires approximately 130GB of memory, which exceeds the capacity of a single HPU (Gaudi 2 with 98GB memory). To address this, we can use the Zero2/Zero3 stages of DeepSpeed (model parallelism) to reduce the memory requirements.
+
+Our tests have shown that training this model requires at least four HPUs when using DeepSpeed Zero2.
+
+```bash
+python ../../gaudi_spawn.py --world_size 4 --use_deepspeed training_nli.py intfloat/e5-mistral-7b-instruct --deepspeed ds_config.json --bf16 --no-use_hpu_graphs_for_training --learning_rate 1e-7
+```
+In the above command, we need to enable lazy mode with a learning rate of `1e-7` and configure DeepSpeed using the `ds_config.json` file. To further reduce memory usage, change the stage to 3 (DeepSpeed Zero3) in the `ds_config.json` file.
+
+# Dataset
 
 We combine [SNLI](https://huggingface.co/datasets/stanfordnlp/snli) and [MultiNLI](https://huggingface.co/datasets/nyu-mll/multi_nli) into a dataset we call [AllNLI](https://huggingface.co/datasets/sentence-transformers/all-nli). These two datasets contain sentence pairs and one of three labels: entailment, neutral, contradiction:
 
@@ -58,7 +82,7 @@ We combine [SNLI](https://huggingface.co/datasets/stanfordnlp/snli) and [MultiNL
 
 We format AllNLI in a few different subsets, compatible with different loss functions. See [triplet subset of AllNLI](https://huggingface.co/datasets/sentence-transformers/all-nli/viewer/triplet) as example.
 
-## SoftmaxLoss
+# SoftmaxLoss
 
 <img src="https://raw.githubusercontent.com/UKPLab/sentence-transformers/master/docs/img/SBERT_SoftmaxLoss.png" alt="SBERT SoftmaxLoss" width="250"/>
 

examples/sentence-transformers-training/nli/ds_config.json

@@ -0,0 +1,16 @@
+{
+    "steps_per_print": 1,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "gradient_accumulation_steps": "auto",
+    "bf16": {
+        "enabled": true
+    },
+    "gradient_clipping": 1.0,
+    "zero_optimization": {
+        "stage": 2,
+        "overlap_comm": false,
+        "reduce_scatter": false,
+        "contiguous_gradients": false
+    }
+}

examples/sentence-transformers-training/nli/training_nli.py

@@ -4,8 +4,8 @@
 STS benchmark dataset
 """
 
+import argparse
 import logging
-import sys
 from datetime import datetime
 
 from datasets import load_dataset
@@ -28,16 +28,43 @@ def main():
     logging.basicConfig(format="%(asctime)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO)
 
     # You can specify any Hugging Face pre-trained model here, for example, bert-base-uncased, roberta-base, xlm-roberta-base
-    model_name = sys.argv[1] if len(sys.argv) > 1 else "bert-base-uncased"
-    train_batch_size = 16
+    parser = argparse.ArgumentParser()
+    parser.add_argument("model_name", help="model name or path", default="bert-base-uncased", nargs="?")
+    parser.add_argument("--peft", help="use LoRA", action="store_true", default=False)
+    parser.add_argument("--lora_target_modules", nargs="+", default=["query", "key", "value"])
+    parser.add_argument("--bf16", help="use bf16", action="store_true", default=False)
+    parser.add_argument(
+        "--use_hpu_graphs_for_training",
+        help="use hpu graphs for training",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+    )
+    parser.add_argument("--learning_rate", help="learning rate", type=float, default=5e-5)
+    parser.add_argument("--deepspeed", help="deepspeed config file", default=None)
+    parser.add_argument("--train_batch_size", help="train batch size", default=16, type=int)
+    args = parser.parse_args()
 
     output_dir = (
-        "output/training_nli_" + model_name.replace("/", "-") + "-" + datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+        "output/training_nli_" + args.model_name.replace("/", "-") + "-" + datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
     )
 
     # 1. Here we define our SentenceTransformer model. If not already a Sentence Transformer model, it will automatically
     # create one with "mean" pooling.
-    model = SentenceTransformer(model_name)
+    model = SentenceTransformer(args.model_name)
+    if args.peft:
+        from peft import LoraConfig, get_peft_model
+
+        peft_config = LoraConfig(
+            r=16,
+            lora_alpha=64,
+            lora_dropout=0.05,
+            bias="none",
+            inference_mode=False,
+            target_modules=args.lora_target_modules,
+        )
+        model = get_peft_model(model, peft_config)
+        model.print_trainable_parameters()
+        model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
 
     # 2. Load the AllNLI dataset: https://huggingface.co/datasets/sentence-transformers/all-nli
     # We'll start with 10k training samples, but you can increase this to get a stronger model
@@ -66,16 +93,16 @@ def main():
     dev_evaluator(model)
 
     # 5. Define the training arguments
-    args = SentenceTransformerGaudiTrainingArguments(
+    stargs = SentenceTransformerGaudiTrainingArguments(
         # Required parameter:
         output_dir=output_dir,
         # Optional training parameters:
         num_train_epochs=1,
-        per_device_train_batch_size=train_batch_size,
-        per_device_eval_batch_size=train_batch_size,
+        per_device_train_batch_size=args.train_batch_size,
+        per_device_eval_batch_size=args.train_batch_size,
         warmup_ratio=0.1,
         # fp16=True,  # Set to False if you get an error that your GPU can't run on FP16
-        # bf16=False,  # Set to True if you have a GPU that supports BF16
+        bf16=args.bf16,  # Set to True if you have a GPU that supports BF16
         # Optional tracking/debugging parameters:
         evaluation_strategy="steps",
         eval_steps=100,
@@ -87,16 +114,18 @@ def main():
         use_habana=True,
         gaudi_config_name="Habana/bert-base-uncased",
         use_lazy_mode=True,
-        use_hpu_graphs=True,
+        use_hpu_graphs=args.use_hpu_graphs_for_training,
         use_hpu_graphs_for_inference=False,
-        use_hpu_graphs_for_training=True,
+        use_hpu_graphs_for_training=args.use_hpu_graphs_for_training,
         dataloader_drop_last=True,
+        learning_rate=args.learning_rate,
+        deepspeed=args.deepspeed,
     )
 
     # 6. Create the trainer & start training
     trainer = SentenceTransformerGaudiTrainer(
         model=model,
-        args=args,
+        args=stargs,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
         loss=train_loss,
@@ -119,6 +148,11 @@ def main():
     final_output_dir = f"{output_dir}/final"
     model.save(final_output_dir)
 
+    if args.peft:
+        model.eval()
+        model = model.merge_and_unload()
+        model.save_pretrained(f"{output_dir}/merged")
+
 
 if __name__ == "__main__":
     main()

examples/sentence-transformers-training/sts/README.md

@@ -5,6 +5,8 @@ Semantic Textual Similarity (STS) assigns a score on the similarity of two texts
 - **[training_stsbenchmark.py](training_stsbenchmark.py)** - This example shows how to create a SentenceTransformer model from scratch by using a pre-trained transformer model (e.g. [`distilbert-base-uncased`](https://huggingface.co/distilbert/distilbert-base-uncased)) together with a pooling layer.
 - **[training_stsbenchmark_continue_training.py](training_stsbenchmark_continue_training.py)** - This example shows how to continue training on STS data for a previously created & trained SentenceTransformer model (e.g. [`all-mpnet-base-v2`](https://huggingface.co/sentence-transformers/all-mpnet-base-v2)).
 
+# General Models
+
 ## Single-card Training
 
 To fine tune on the STS task:
@@ -33,7 +35,30 @@ For multi-card training you can use the script of [gaudi_spawn.py](https://githu
 HABANA_VISIBLE_MODULES="2,3" python ../../gaudi_spawn.py --use_deepspeed --world_size 2 training_stsbenchmark.py bert-base-uncased
 ```
 
-## Training data
+
+# Large Models (intfloat/e5-mistral-7b-instruct Model)
+
+## Single-card Training with LoRA+gradient_checkpointing
+
+Pretraining the `intfloat/e5-mistral-7b-instruct` model requires approximately 130GB of memory, which exceeds the capacity of a single HPU (Gaudi 2 with 98GB memory). To address this, we can utilize LoRA and gradient checkpointing techniques to reduce the memory requirements, making it feasible to train the model on a single HPU.
+
+```bash
+python training_stsbenchmark.py intfloat/e5-mistral-7b-instruct --peft --lora_target_modules "q_proj" "k_proj" "v_proj"
+```
+
+## Multi-card Training with Deepspeed Zero2/3
+
+Pretraining the `intfloat/e5-mistral-7b-instruct` model requires approximately 130GB of memory, which exceeds the capacity of a single HPU (Gaudi 2 with 98GB memory). To address this, we can use the Zero2/Zero3 stages of DeepSpeed (model parallelism) to reduce the memory requirements.
+
+Our tests have shown that training this model requires at least four HPUs when using DeepSpeed Zero2.
+
+```bash
+python ../../gaudi_spawn.py --world_size 4 --use_deepspeed training_stsbenchmark.py intfloat/e5-mistral-7b-instruct --deepspeed ds_config.json --bf16 --no-use_hpu_graphs_for_training --learning_rate 1e-7
+```
+
+In the above command, we need to enable lazy mode with a learning rate of `1e-7` and configure DeepSpeed using the `ds_config.json` file. To further reduce memory usage, change the stage to 3 (DeepSpeed Zero3) in the `ds_config.json` file.
+
+# Training data
 
 Here is a simplified version of our training data:
 
@@ -70,7 +95,7 @@ train_dataset = load_dataset("sentence-transformers/stsb", split="train")
 # })
 ```
 
-## Loss Function
+# Loss Function
 
 <img src="https://raw.githubusercontent.com/UKPLab/sentence-transformers/master/docs/img/SBERT_Siamese_Network.png" alt="SBERT Siamese Network Architecture" width="250"/>
 

examples/sentence-transformers-training/sts/ds_config.json

@@ -0,0 +1,16 @@
+{
+    "steps_per_print": 1,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "gradient_accumulation_steps": "auto",
+    "bf16": {
+        "enabled": true
+    },
+    "gradient_clipping": 1.0,
+    "zero_optimization": {
+        "stage": 2,
+        "overlap_comm": false,
+        "reduce_scatter": false,
+        "contiguous_gradients": false
+    }
+}

examples/sentence-transformers-training/sts/training_stsbenchmark.py

@@ -4,8 +4,8 @@
 
 """
 
+import argparse
 import logging
-import sys
 from datetime import datetime
 
 from datasets import load_dataset
@@ -25,19 +25,48 @@ def main():
     logging.basicConfig(format="%(asctime)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO)
 
     # You can specify any Hugging Face pre-trained model here, for example, bert-base-uncased, roberta-base, xlm-roberta-base
-    model_name = sys.argv[1] if len(sys.argv) > 1 else "distilbert-base-uncased"
+    parser = argparse.ArgumentParser()
+    parser.add_argument("model_name", help="model name or path", default="distilbert-base-uncased", nargs="?")
+    parser.add_argument("--peft", help="use LoRA", action="store_true", default=False)
+    parser.add_argument("--lora_target_modules", nargs="+", default=["q_lin", "k_lin", "v_lin"])
+    parser.add_argument("--bf16", help="use bf16", action="store_true", default=False)
+    parser.add_argument(
+        "--use_hpu_graphs_for_training",
+        help="use hpu graphs for training",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+    )
+    parser.add_argument("--learning_rate", help="learning rate", type=float, default=5e-5)
+    parser.add_argument("--deepspeed", help="deepspeed config file", default=None)
+    args = parser.parse_args()
+
     train_batch_size = 16
     num_epochs = 1
     output_dir = (
         "output/training_stsbenchmark_"
-        + model_name.replace("/", "-")
+        + args.model_name.replace("/", "-")
         + "-"
         + datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
     )
 
     # 1. Here we define our SentenceTransformer model. If not already a Sentence Transformer model, it will automatically
     # create one with "mean" pooling.
-    model = SentenceTransformer(model_name)
+    model = SentenceTransformer(args.model_name)
+
+    if args.peft:
+        from peft import LoraConfig, get_peft_model
+
+        peft_config = LoraConfig(
+            r=16,
+            lora_alpha=64,
+            lora_dropout=0.05,
+            bias="none",
+            inference_mode=False,
+            target_modules=args.lora_target_modules,
+        )
+        model = get_peft_model(model, peft_config)
+        model.print_trainable_parameters()
+        model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
 
     # 2. Load the STSB dataset: https://huggingface.co/datasets/sentence-transformers/stsb
     train_dataset = load_dataset("sentence-transformers/stsb", split="train")
@@ -61,7 +90,7 @@ def main():
     )
 
     # 5. Define the training arguments
-    args = SentenceTransformerGaudiTrainingArguments(
+    stargs = SentenceTransformerGaudiTrainingArguments(
         # Required parameter:
         output_dir=output_dir,
         # Optional training parameters:
@@ -70,7 +99,7 @@ def main():
         per_device_eval_batch_size=train_batch_size,
         warmup_ratio=0.1,
         # fp16=True,  # Set to False if you get an error that your GPU can't run on FP16
-        # bf16=True,  # Set to True if you have a GPU that supports BF16
+        bf16=args.bf16,  # Set to True if you have a GPU that supports BF16
         # Optional tracking/debugging parameters:
         evaluation_strategy="steps",
         eval_steps=100,
@@ -82,16 +111,18 @@ def main():
         use_habana=True,
         gaudi_config_name="Habana/distilbert-base-uncased",
         use_lazy_mode=True,
-        use_hpu_graphs=True,
+        use_hpu_graphs=args.use_hpu_graphs_for_training,
         use_hpu_graphs_for_inference=False,
-        use_hpu_graphs_for_training=True,
+        use_hpu_graphs_for_training=args.use_hpu_graphs_for_training,
+        learning_rate=args.learning_rate,
+        deepspeed=args.deepspeed,
     )
 
     # 6. Create the trainer & start training
     # trainer = SentenceTransformerTrainer(
     trainer = SentenceTransformerGaudiTrainer(
         model=model,
-        args=args,
+        args=stargs,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
         loss=train_loss,
@@ -113,6 +144,11 @@ def main():
     final_output_dir = f"{output_dir}/final"
     model.save(final_output_dir)
 
+    if args.peft:
+        model.eval()
+        model = model.merge_and_unload()
+        model.save_pretrained(f"{output_dir}/merged")
+
 
 if __name__ == "__main__":
     main()

optimum/habana/sentence_transformers/__init__.py

@@ -19,5 +19,5 @@
 from .st_gaudi_trainer import SentenceTransformerGaudiTrainer
 from .st_gaudi_training_args import SentenceTransformerGaudiTrainingArguments
 from .st_gaudi_encoder import st_gaudi_encode
-from .st_gaudi_transformer_tokenize import st_gaudi_transformer_tokenize
+from .st_gaudi_transformer import st_gaudi_transformer_tokenize, st_gaudi_transformer_save
 from .st_gaudi_data_collator import st_gaudi_data_collator_call

optimum/habana/sentence_transformers/modeling_utils.py

@@ -25,6 +25,7 @@ def adapt_sentence_transformers_to_gaudi():
     from optimum.habana.sentence_transformers import (
         st_gaudi_data_collator_call,
         st_gaudi_encode,
+        st_gaudi_transformer_save,
         st_gaudi_transformer_tokenize,
     )
 
@@ -33,6 +34,7 @@ def adapt_sentence_transformers_to_gaudi():
     from sentence_transformers.models import Transformer
 
     Transformer.tokenize = st_gaudi_transformer_tokenize
+    Transformer.save = st_gaudi_transformer_save
 
     from sentence_transformers.data_collator import SentenceTransformerDataCollator